Add causal LoRA adapter lifecycle

[WaldenLee2005]

Summary

• Add strict LoRA adapter continuation with --resume-from and intentional training forks with --branch-from.
• Add validated adapter staging for local and cloud artifacts, SHA-256 attestation, parent lineage, and recorded training recipes.
• Add yeto merge for safe local adapter merges into an immutable base model with atomic, sharded SafeTensors output.
• Document lifecycle semantics, safety checks, limitations, and operator workflows.

Behavior and safety

• Strict resume requires the recorded base model, dataset identity, trust setting, and training recipe to match.
• Branching permits deliberate recipe or data changes while preserving parent provenance.
• Adapter continuation restores trainable LoRA weights but starts a fresh optimizer and scheduler; syncer checkpoints remain the crash-recovery mechanism.
• Parent adapters must be SafeTensors LoRA artifacts for causal language models and must pass base-model, rank, alpha, digest, and symlink checks.
• Cloud adapters require an explicit expected SHA-256 digest before GPU work begins.
• Merge uses merge_and_unload(safe_merge=True), publishes atomically, records provenance, and supports configurable output shard size. Output sharding limits file size, not peak merge memory.

Validation

Automated

• Post-rebase focused regression: 115 passed.
• Pre-rebase non-MLX repository regression: 776 passed and 16 skipped; two log-sensitive syncer assertions also passed separately with the expected RUST_LOG=info setting.
• Local MLX collection remains blocked by an unrelated fatal MLX import in the Mac environment.

H200 manual validation

• SmolLM2 135M NF4 QLoRA: initial training, branch, strict resume, intentional drift rejection, tensor-change verification, merge, 24-shard SafeTensors export, provenance checks, reload, and generation all passed.
• Qwen3.6 27B at revision 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9: initial NF4 QLoRA, branch from sequence length 128 to 256, strict resume, lineage, all 992 adapter tensors changing, safe bf16 merge, 12 SafeTensors shards totaling 50.1 GiB, provenance, and merged-weight reload all passed on one H200.
• Cached 27B generation exposed an incompatible optional causal_conv1d runtime fast path, not a merge or weight failure. Generation passed both with cache disabled and with the Transformers torch causal-convolution fallback forced. The production serving runtime should still be validated independently.